Siti Norul Huda Sheikh Abdullah

mimi@ftsm.ukm.my

� 90 % is � 90 % is

standard

license plates

� 10% is in

special

format

� Definitions: Car plate recognition, plate number recognition, vision plate, automatic number plate recognition and etc. (Hofman, Y., 2004)

� Most of the research has shown that classification rate was higher than detection rate. Therefore, license plate detection is a crucial research because it correlates with problem statements.

� Some research have been conducted world wide.

YearYear DesignerDesigner License formatLicense format TechniquesTechniques AccuracyAccuracyYearYear DesignerDesigner License formatLicense format TechniquesTechniques AccuracyAccuracy

19981998 Hans. A. Herg et. al

DutchDutch Histogram, Hotelling Histogram, Hotelling TransformationTransformation

98.7 %98.7 %

19991999 T.Naito et..al. JapaneseJapanese Templat matchingTemplat matching 97%97%

20032003 Shen-Zen Whang and His Jian Lee

TaiwaneseTaiwanese Sobel, disciriminant Sobel, disciriminant functionfunction

96.60%96.60%

20032003 M.Safraz et.al ArabicArabic Sobel, templat matchingSobel, templat matching 95.24%95.24%

20042004 TomohikoNukano et.al

MalaysianMalaysian NN ThresholdNN Threshold 87.3%87.3%

� | Hcurrent – H’| <= (U+V) x Hcurrent

� | minYcurrent –Y’ | <= (U+V) x Hcurrent

threshold average

fixed (130) manual

total sample data 1000 1000 1000

no of correct 803 919 861

correct percentage 80.3 % 91.9 % 86.%

segmentation error 121 10 65.5

Abdullah, S. N. H. S.; Khalid, M.; Yusof, R. & Omar, K.License Plate Recognition Using Multi-classifier and Neural Network 2nd IEEE International Conference on Information & Communication Technologies: from Theory to Applications (ICTTA'06), 2006, 659-660. (CD Proceedings)

Segmentation error percentage 61.4 % 12.3 % 36.9 %

classification error 76 71 73.5

Classification error percentage 38.6 % 87.7 % 63.1 %

Why ?1.Connected

characters2. Symbol or sign on

LP.3. Cannot distinguish

slanting 30

40

50

60

Num

ber

31

slanting characters, fails to detect similar look characters. Ex: 3->8, B->8, 6->G or 4->A etc.

4. This problem occurs on 2 line LP.

0

10

20

30

Not

Fou

nd

Mis

s_1

Mis

s_2

Mis

s_>2

Ext

ra

Wro

ng_1

Wro

ng_2

Wro

ng_>

2

Wro

ng_s

eq

Type

Num

ber

f ixed (130)

dif ferent

42

� Illumination often distracts determining object from the source image (Petrou and Bosdogianni, 2000)

� Improper segmentation technique contributes reduction in recognition rates (Lee et al.,2004).

� Otsu Thresholding technique applied statistical � Otsu Thresholding technique applied statistical approach in calculating acceptable threshold value based on histogram projection. It only suitable in solving threshold localization process which makes it less accurate on globalization process.

� Therefore, an alternative method is required to improve the segmentation process.

1. To develop an enhance license plate detection (LPD),

2. To develop an alternative feature extraction and compare it with different classification and compare it with different classification techniques

3. To determine adaptive threshold value for image segmentation.

Hence, this research consists of three contributions:

First contributionObjective: To develop an enhance license plate detection (LPD)

Abdullah, S. N. H. S.; Khalid, M.; Yusof, R. & Omar, K.Cluster Run Length Smoothing Algorithm (CRLSA) for License Plate Detection Proceedings of the 25th International Multi-Conference on Artifical Intelligence and Applications (IASTED 2007), the 25th International Multi-Conference on Artifical Intelligence and Applications (IASTED 2007), 2007, 323-328

Abdullah, S. N. H. S.; Khalid, M.; Yusof, R. & Zainal, N.F.A. 2009. License plate detection and segmentation using cluster run length smoothing algorithm. IEEE Potential Journal. (Accepted for publish)

� Run Length Smoothing Algorithm (RLSA) has been used widely in optical character recognition (OCR) process especially in document analysis [Strouthopoulos et al. 1999].

� Normally RLSA is only applied after horizontal or � Normally RLSA is only applied after horizontal or vertical projection to recognize the block segmentation and the transformation are only concerned on two dimensional image (black − white) [Gatos & Papamarkos, 2001]

Strouthopoulos, C.; Papamarkos, N. & Chamzas, C.PLA using RLSA and a neural networkEngineering Applications of Artificial Intelligence, 1999, 12, 119-138

Gatos, B. & Papamarkos, N.Applying fast segmentation techniques at a binary image represented by a set of non-overlapping blocksProceedings. Sixth International Conference on Document Analysis and Recognition, 2001., 2001, 1147 - 1151

[5] Siah, Y. K. A design of an intelligent license plate recognition. Universiti Teknologi Malaysia, 2000[23] Saldago, L.; Menende, J.; Rendon, E. & Garcia., N. Automatic car plate recognition through vision

engineering.Proceedings of IEEE 33rd Annual 1999 International Carnahan Conference on security technology, 1999, 71-76.

[1] Han, P.; Han, W.; Wang, D. & Zhai, Y. Car License Plate feature extraction and recognition based on multistage classifier.International conference on Machine Learning and Cybernetics, 2003, 1, 128-13

[2] Miyamoto, K.; Nagano, K. & Tamagawa, M. Vehicle license-plate recognition by image analysisProceedings of IEEE International Conference on Industrial electronics, Control and Instrumentation, 1991, 3, 1734-1738

[22] Kong, J.; Liu, X.; Lu, Y.; Zhou, X. & Zhao, Q. Zhang, S. & Jarvis, R. (ed.) Robust License Plate Segmentation Method Based on Texture Features and Radon Transform AI 2005: Advances in Artificial Intelligence Springer-Verlag Berlin Heidelberg, 2005, 3809/2005, 510-519

The proposed

technique where:

� The CT sub-module

comprises three

main processes:

(a) Run RGB

Convolution (RGBC)

and Check the First

Run Length

.

303

503

303

128=),(

303

505

303

128=),(

−

−

−

+

−

−

−

+

verticalKED

verticalPED

yxG

yxG

,=0 p

RunLengthaveRLwhole

p

∑

,=0 p

ngthCountRunLewholeaveCountRL

p

∑

.=cCHp

.= HLongtotalHLongq

∑Run Length

Smoothing

Algorithm (RLSA1),

(b) Run Threshold Value

One (TV1) and Check

the Second RLSA

(RLSA2),

(c) Get winner cluster

indexes.

.=0

HLongtotalHLong ∑

,=0

HMediumumtotalHMediq

∑

.=0

HShortttotalHShorq

∑

� Benefits:� Even though the original image of the back or

frontalcar is having fusion problems such as illumination, blur, rotated, skewed, CRLSA can still successfully detect the locations of the license plate.

� CRLSA is able to detect the precise location of the license plate even though the license plates’scharacters are connected or too close to each other.characters are connected or too close to each other.

� Limitations:� Processing time. RGB convolution with a edge

detector is time consuming� Memory allocation. Quite often memory leaks when

using RGB convolution because it requires high storage.

� Future Works:� Instead of using RGB Convolution using individual

pixels, why not apply feature vectors. � Introduce new algorithm for adaptive thresholding.

Second contributionObjective 2: To develop an alternative feature extraction and compare it with different classification techniques

Third contribution

Objective 3:To determine adaptive threshold value for image segmentation.

� Thresholding is a straightforward technique in transforming

a gray scale image into a binary image that can facilitate the

segmentation process. If , I is the input image, the value of

the output image, , at position (x, y), given a threshold , is:

� Other methods for calculating the threshold namely the method of local entropy, proposed by Shanon [10] and Otsu[11].

� Otsu Thresholding is a favourite optimization thresholdingmethod (Siah, 2000; Shapiro et al., 2006; Kahraman et al., 2003). It adopts primitive mean, μ, (Equation 2.8a) and variance, σ2, (Equation 2.8b) calculations from histogram distributions , ρi, of an (Equation 2.8b) calculations from histogram distributions , ρi, of an image, ρ where i is from 0 until g of gray level values.

[10] Yan, C.; Sang, N. & Zhang, T. Local entropy-based transition region extraction and thresholding Pattern Recognition Letters, 2003, 24, 2935 - 2941

[11] Noboyuki Otsu, 1979. A Threshold Selection Method from Gray-Level Histograms, IEEE Transactions on System, Man and Cybernetics, Vol SMC-9 (1):62-66.

The proposed threshold (PT) consists of three

steps:

(1) Identify the type of image according to its

histogram, histogram,

(2) Calculation of blob distributions for various

threshold values

(3) Selection of thresholds.

1

Identify the

type of image

according to

its histogramits histogram

LPR+PT

Fixed Threshold Threshold value

Otsu Threshold value

� Advantages :� The proposed framework has significantly increased the LPD and LPS

accuracy rate and it can give an alternative way to the LPR system.

� The proposed threshold gives an alternative solution to obtain better threshold value(s) based on rule of thumb.

� The proposed threshold equally treats all global and local maximum � The proposed threshold equally treats all global and local maximum values of the probability density. Unlike Otsu, which only selects the global maximum value as the best threshold value which might not necessarily be true.

� Disadvantages� The proposed framework is prone to detect unnecessary blobs and

signs together with the license plate characters.

� Still not robust enough for real time applications due to overall LPR processing time because it suggests a selection of threshold values.

� Apply two dimensional blob distributions and

apply mean and standard deviations to select

only one threshold values.

Apply local entropy to select the best peaks � Apply local entropy to select the best peaks

of the blob distributions.